🤖 AI

Can Causal Discovery Algorithms Help in Generating Legal Arguments?

This paper demonstrates that causal discovery algorithms, when applied to a curated dataset of 150 annotated homicide cases, can successfully uncover probabilistic causal relationships between legal concepts to support the automated generation of viable legal arguments.

Original authors: Soham Wasmatkar, Subinay Adhikary, Rakshit Rohan, Shouvik Kumar Guha, Saptarshi Pyne, Kripabandhu Ghosh

Published 2026-05-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Soham Wasmatkar, Subinay Adhikary, Rakshit Rohan, Shouvik Kumar Guha, Saptarshi Pyne, Kripabandhu Ghosh

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a massive puzzle where the pieces are thousands of old court cases. Usually, a human lawyer has to read through these cases one by one, looking for patterns like "When a fight happens over land, it often leads to a murder." This is slow, and humans can only look at a few hundred cases at a time.

This paper asks a simple question: Can we teach a computer to find these hidden patterns automatically, and can those patterns help lawyers build better arguments?

Here is how the authors tried to answer that, explained in everyday terms:

1. The "Detective's Notebook" (The Dataset)

The researchers didn't just throw random text at a computer. They acted like careful librarians.

They picked 150 real-life murder cases from court records.
They created a list of 17 specific "concepts" (like "Physical Assault," "Property Dispute," "Riot," or "Death Sentence").
They went through every single case and marked it with these concepts. For example, if a case involved a fight with a knife, they marked it "Physical Assault." If it was about a fight over land, they marked it "Property Dispute."
Think of this as turning 150 messy stories into a neat spreadsheet of Yes/No answers for each concept.

2. The "Pattern Hunters" (The Algorithms)

Next, they didn't just use one tool; they used six different "pattern-hunting" algorithms. You can think of these as six different detectives, each with a unique style of investigation:

Some detectives look for things that happen together.
Some try to figure out which thing caused the other (e.g., did the riot cause the murder, or did the murder cause the riot?).
They fed the 150 cases into these six detectives and asked: "What causes what?"

3. The "Consensus" (The Results)

Since the six detectives had different styles, they didn't all agree on everything. To find the truth, the researchers looked for the consensus—the patterns that at least half of the detectives agreed on.

They found some very interesting connections that a simple "counting" method would have missed:

The "Political Rivalry" Clue: The algorithms discovered that if a case had no consistent evidence AND involved a riot, it was almost 100% certain to be connected to political rivalry.
- Why this matters: A simple count might say, "Riots happen in many cases, so they aren't special." But the computer saw that specifically when evidence is missing AND a riot happens, the political motive is a lock.
The "No Fight" Clue: This is the paper's biggest "aha!" moment. The algorithms found a strong rule: If there was NO physical assault in a murder case, it is 100% certain that the murder was NOT caused by a property dispute.
- The Analogy: Imagine you are trying to prove someone stole a car because they wanted to sell the parts. If you can prove the car was never even touched or damaged (no physical assault), you can instantly rule out the "selling parts" motive. The computer found this logic hidden in the data.

4. Why This is Different from Old Methods

The paper explains that old ways of analyzing data (like looking for simple correlations) are like looking at a blurry photo. They might tell you, "Riots and politics are somewhat related," but they miss the specific details.

These new algorithms are like a high-definition microscope. They don't just say "A and B are related." They say, "If A happens and C happens, then B is guaranteed." This allows lawyers to make arguments with mathematical certainty (e.g., "This proves the motive is X with 100% probability").

5. The Bottom Line

The paper concludes that yes, these algorithms can help generate legal arguments.

They can turn a messy pile of 150 cases into a clear map of cause-and-effect.
They can provide quantitative arguments (numbers and probabilities) that are fully explainable.
The authors suggest a future workflow where the computer finds the patterns, and a human lawyer checks them to make sure they make sense legally.

What the paper does NOT claim:

It does not say these algorithms can replace judges or lawyers.
It does not claim they work on all types of law (they only tested murder cases).
It does not say they can predict the future or decide who is guilty in a new case instantly.
It strictly focuses on using these tools to generate arguments based on past data, not to make final legal rulings.

In short, the paper shows that computers can act as powerful assistants, helping lawyers spot hidden logical connections in past cases that humans might miss, making legal arguments stronger and more data-driven.

Technical Summary: Causal Discovery Algorithms for Legal Argument Generation

Problem Statement
While causal discovery algorithms have become standard tools in fields like medicine and economics for analyzing large multivariate datasets and uncovering causal relationships, their application in the legal domain remains unexplored. Current legal literature relies on causal relationships manually identified by legal experts, which limits the scale of analysis and lacks data-driven quantification of belief in these relationships. This paper addresses the gap by investigating whether automated causal discovery algorithms can be leveraged to generate viable, quantitative legal arguments from historical case data.

Methodology
The study employs a structured pipeline involving dataset preparation, algorithmic application, and consensus building:

Dataset Construction: The authors curated a novel dataset of 150 homicide cases from Indian court records. These cases were annotated with 17 distinct legal concepts (e.g., "physical assault," "property dispute," "riot," "death sentence"). The annotation process involved three legal experts using the LeDA tool, with Inter-Annotator Agreement (IAA) measured via F1-score. Cases with IAA scores below 0.8 were reviewed by a senior expert. The final output was a 150-by-17 binary matrix encoding the presence (1) or absence (0) of each concept per case.
Algorithm Selection: Six widely-used causal discovery algorithms were selected based on methodological diversity:
- PC: Builds graphs via conditional independence tests.
- GES: Uses score-based local optimization to add/remove edges.
- GRaSP & BOSS: Order-based methods that search for optimal variable orderings to construct graphs.
- LiNGAM & ANM: Regression-based methods (linear and non-linear, respectively) modeling concepts as target/predictor variables.
Consensus Formation: To ensure robustness, a "consensus causal graph" was constructed by identifying relationships discovered by more than one algorithm. Relationships were weighted by the number of algorithms agreeing on them.
Quantification: The strength of discovered relationships was quantified using Conditional Probability Tables (CPTs), allowing for the calculation of specific probabilities (e.g., $P(\text{Effect} | \text{Cause})$ ).

Key Results
The application of these algorithms yielded a consensus causal graph containing six robust relationships, agreed upon by at least half of the algorithms. Key findings include:

Political Rivalry Prediction: Three algorithms (PC, GRaSP, BOSS) identified that a combination of "evidence inconsistency" and "riot" serves as a predictive cause for "political rivalry" with a probability of 1.0. Notably, standard Pearson correlation failed to detect this strong association (reporting low correlations of ~0.36–0.39) because it aggregates over value combinations rather than analyzing specific joint states.
Physical Assault and Property Dispute: A consensus was reached that the absence of a physical assault is a sufficient condition (probability = 1.0) to establish that a homicide was not committed due to a property dispute. Conversely, the presence of physical assault only marginally increases the probability of a property dispute motive (from 0.00 to 0.07). This relationship, with a low Pearson correlation of 0.2, was undetectable via traditional correlation measures but clearly identified by causal discovery.
Argument Generation: The study demonstrates that these algorithms can generate quantitative arguments. For instance, establishing that a defendant is not an heir to a victim's wealth was shown to increase the probability of proving innocence by 59% in a hypothetical scenario, derived directly from the CPTs.

Significance and Claims
The paper claims that causal discovery algorithms can effectively assist in generating legal arguments that are both legally viable and fully explainable in mathematical and legal terms. The significance of this work lies in:

Scalability and Quantification: Unlike manual expert analysis, these algorithms can process vast numbers of prior cases to quantify the degree of belief in causal links, offering a data-driven basis for argumentation.
Beyond Correlation: The study highlights that causal discovery captures fine-grained, directed associations between specific value combinations of variables that traditional correlation measures miss.
Transparency and Justice: The authors posit that providing transparent causal structures helps litigants understand the reasoning behind judgments. This transparency is crucial for evaluating whether outcomes reflect the severity of the crime or procedural factors (e.g., prosecutorial delay).
Future Potential: While the current work focuses on "association-based reasoning" (the lowest level of the ladder of causation), the paper suggests that integrating these algorithms with higher-level causal reasoning (such as counterfactuals) could further enhance legal analysis. The authors envision a pipeline where automated discovery is refined by human legal experts to improve efficiency and reduce false discoveries, representing a conceptual advancement toward timely justice and transparency in law.

The paper concludes modestly, noting that while the current dataset is limited, the framework opens avenues for future endeavors to maximize the utilization of causal discovery in automated legal argument generation.